ACBD3 claim-consistency audit

audit_acbd3_claims.py is a regression test over this review's claims, not its quotes.

Why it exists

checkquotes.py and the repo's reference validator verify that each
supporting_text is verbatim in its source. Nothing verified that a claim is
stated consistently at the five independent sites that can assert it:

None is generated from the others. Over PR #2232, revising a claim at the flagged
site while leaving it stale elsewhere accounted for review rounds 10, 12, 13, 14
and 15 — after the biology had settled at round 9.

What it checks

• 29 retracted phrasings must not appear in either of the two files the
  review authors: ACBD3-ai-review.yaml and ACBD3-notes.md. It deliberately does
  not scan ACBD3-deep-research-affinage.md — that is a machine-fetched provider
  record which must not be edited, and FORBIDDEN[0] (PI4KB "through its GOLD
  domain") is still live in it, correctly so. The audit governs what this review
  asserts, not what the provider said.
• A retracted phrasing may legitimately be quoted inside its own retraction, so a
  match is suppressed when a retraction marker falls in a window weighted backwards
  (-170/+140 characters): retractions normally introduce the phrasing they retract,
  but some place the marker just after the quote, hence the smaller forward allowance.
• Matching strips markdown emphasis as well as collapsing whitespace. Without
  that, dispensable for **3A-mediated** PI4KB recruitment in the notes — the most
  explicit statement of that scope anywhere in the review — counted for nothing.
• 9 claims must remain positively present, each with a minimum occurrence count.
  This catches what a forbidden-string list cannot: a regression that removes a
  qualifier rather than adding a wrong phrase. 3A-mediated PI4KB recruitment is
  the case that motivated it — the scoped form was written in c1ba137, deleted by
  the round-9 rewrite in 81f46c582, and the loss was invisible to every other check.

The count is load-bearing, and the first version of this file did not have it.
An "appears anywhere" check cannot see a claim removed from one of several sites
— precisely the failure it was written to prevent. The self-test caught that:
deleting the qualifier from one of its two sites still reported zero problems.

Every floor equals the current actual count, with zero slack. A floor set even
slightly low silently restores the hole it was meant to close — 3A-mediated PI4KB recruitment was first committed at 5 against an actual 8, leaving room for
three assertions of the scope to be deleted unnoticed. Two of the eight are
supporting_text quotes that checkquotes.py already protects, so the effective
prose guard was weaker still. If a claim is legitimately restated more often later,
raise the floor deliberately rather than leaving headroom.
- Matching is whitespace-normalised, so a phrase broken across a line wrap is
still found. Several regressions survived plain grep for exactly this reason.

Self-test

The window heuristic is the fragile part: written first as a symmetric 320-character
window, it silently suppressed the very regression being hunted. So the intended
verification is:

uv run python genes/human/ACBD3/ACBD3-bioinformatics/audit_acbd3_claims.py   # expect 0 problems
# then break it deliberately, three ways, and confirm each is caught:
#   1. reintroduce a retracted phrasing            (round-10 sentence in the notes)
#   2. delete a scope qualifier from ONE of its sites
#   3. remove the emphasis markers from the notes' bolded scope statement
# restore, confirm clean again

A checker that reports zero problems on a file with a known regression is worse than
no checker — and this script has hit that state three times, each caught only by
trying to break it: the over-wide retraction window, presence-instead-of-count, and a
floor set below the actual count. The script resolves the repository root from its own
path, so it runs from any working directory.

Result on the current tree

29 retracted phrasings checked, 9 required claims checked, 0 problem(s)

The counts in this document are the code's: 29 FORBIDDEN, 9 REQUIRED.

Run against the round-13 tree it found exactly one regression — ACBD3-notes.md
routing all Golgi scaffolding through two domain surfaces, three lines above a
domain map in the same file listing three. That was independently flagged by the
reviewer in the next round.

Scope and honesty

This is a per-gene, hand-maintained lexicon, not a general lint. It cannot find
a claim nobody has yet noticed is wrong; it only stops a known correction from
being undone. Generalising it would mean a lint over the four YAML text slots plus
the notes for any gene whose review has been revised — worth doing, but out of
scope for one gene review.